Deep4ge: Mutation-Induced Training Dynamics Dataset for DNN Fault Detection and Diagnosis

Deep4ge v1.0.0

A dataset of 14,227 mutation-induced per-epoch training logs generated from 60 real-world DNN programs, engineered for fault detection and diagnosis research in deep neural networks.

Dataset contents

• 9,845 faulty training logs across 7 fault categories
• 4,382 correct baseline logs
• 60 StackOverflow seed programs (FNN / CNN / RNN)
• 31 dynamic features per epoch (gradients, activations, weights, hardware)
• Manifest index, data dictionary, and validation scripts

Licenses

• Dataset: CC BY 4.0
• Mutation framework: MIT

Citation

If you use Deep4ge, please cite this release and refer to the CITATION.cff file in the repository.

Submitted to ICST 2026 Testing Tools and Data Showcase Track.